Title：Abnormal Behavior Recognition of Dairy Cows via DeepLabCut-Based Limb Decomposition  

Author：Yan Zhao and Xiaoxu PEI

Journal：IEEE Access（SCI）

Abstract：The early recognition of abnormal behavior in dairy cows is crucial for lameness warnings，disease detection， and welfare assessment. However， traditional methods that rely on holistic features struggleto distinguish normal from abnormal behaviors with similar postures. This paper proposes a hierarchical method spanning keypoint detection， limb decomposition， and abnormal behavior recognition based on the DeepLabCut framework. On a self-constructed 18-keypoint cattle pose dataset of approximately 4，000 annotated frames， two scene-separated Bayesian-optimization (BO) pose networks—ResNet101+BO for single-cow frames and top-down ResNet101+BO for multi-cow frames—serve as skeleton sources with inference-time adaptive switching. The 18 keypoints are organized into five functional limb regions. Geometric features (joint angles， normalized limb lengths， and region centroids) and motion features (velocity， gait cycle， limb coordination， and gait symmetry) are extracted to construct a 40-dimensional structured feature vector. An abnormal behavior detection framework that integrates statistical thresholds and machine learning automatically detects lameness， gait abnormalities， and feeding anomalies. The experimental results demonstrate that the method achieves a behavior recognition accuracy of 91.6% with Extreme Gradient Boosting (XGBoost)， an abnormal gait detection sensitivity of 86.4%， and a specificity of 93.1%. The system achieves a tracking success rate of 96.5% and 28 frames per second (FPS) on a real farming video. This study bridges pose estimation and abnormal behavior understanding， providing an automated technical route for early lameness warnings and health management of dairy cows.

Keywords：Cows，Limbs，Training，Behavior recognition，Skeleton，Modeling，Signal detection，Timing，Head，Frequency